Distributionally Robust Optimization and Invariant Representation Learning for Addressing Subgroup Underrepresentation: Mechanisms and Limitations

Nilesh Kumar, Ruby Shrestha, Zhiyuan Li, Linwei Wang

Introduction

Deep Neural Networks (DNNs) have demonstrated remarkable performance in medical imaging tasks, yet their tendency to exploit "shortcuts"—spurious correlations between non-predictive features and target labels—remains a significant barrier to clinical deployment. In dermatology, for instance, a model might erroneously associate the presence of a surgical bandage with malignancy rather than identifying the morphological characteristics of the lesion itself. This phenomenon is often rooted in subgroup underrepresentation: when specific combinations of features (e.g., malignant lesions without bandages) are rare in the training data, the model prioritizes the more frequent, easier-to-learn shortcut.

Addressing Subgroup Underrepresentation

Current strategies to mitigate these biases generally fall into two categories: Distributionally Robust Optimization (DRO) and Invariant Representation Learning. DRO methods, such as Group DRO (GDRO), aim to minimize the maximum loss across predefined subgroups, effectively upweighting underrepresented or poorly performing groups. In contrast, Invariant Representation Learning, exemplified by Domain Adversarial Neural Networks (DANN), attempts to strip the model’s internal embeddings of sensitive or domain-specific information (e.g., skin tone).

Despite their theoretical appeal, these methods often struggle in complex, real-world scenarios. This research investigates the mechanisms behind these failures and proposes a synergistic framework to improve robustness and fairness in medical diagnostics.

Proposed Methodology: A Hybrid Approach

The authors identify a fundamental limitation in standalone DANN: the primary classifier often "overpowers" the domain-adversarial component, retaining shortcut information because it provides a strong signal for the main task. To address this, they propose a hybrid architecture that integrates GDRO with DANN.

In this framework, the main classifier is optimized using GDRO to ensure robustness across subgroups. Simultaneously, a DANN-based adversarial loss is applied to the feature extractor to enforce representation invariance. By reducing the classifier's reliance on shortcuts through GDRO, the adversarial component can more effectively eliminate spurious features from the embeddings. The process is finalized with Deep Feature Reweighting (DFR), fine-tuning the classification head on a small, balanced validation set to further minimize performance disparities.

Experimental Evaluation

The study evaluated several debiasing techniques, including Generalized Reweighting (GDRO, JTT), Invariant Learning (DANN), and Post-hoc Correction (DFR), across two distinct skin lesion datasets:

  1. ISIC (Artificial Bias): A dataset where a "bandage" shortcut was artificially introduced. Malignant cases without bandages were extremely rare (100 samples), creating a stark spurious correlation.
  2. Fitzpatrick (Natural Bias): A real-world dataset where label distributions (benign vs. malignant) naturally vary across different skin tones (Fitzpatrick scales 1-6), presenting a more nuanced challenge for bias mitigation.

Key Insights into Robustness and Invariance

The research yields several critical insights into why standard debiasing fails:

  • The GDRO "Group Adjustment" Trap: Without specific adjustment terms, GDRO may inadvertently upweight samples that are simply noisy or inherently difficult, rather than those that conflict with the learned bias.
  • The Brittleness of Invariance: When spurious correlations are strong, models are incentivized to "cheat." The adversarial "domain-stripper" in DANN often fails to remove shortcut information if that information is too useful for the primary classification task.
  • Synergistic Effects: The core finding is that robustness and invariance are complementary. Making the classifier robust via GDRO reduces the utility of shortcuts, which facilitates the learning of truly invariant representations by the adversarial component.

Results and Performance

The hybrid approach demonstrated superior performance across both datasets:

  • ISIC Results: The proposed method achieved a worst-group accuracy of 0.800, significantly outperforming the Empirical Risk Minimization (ERM) baseline (0.556) and standalone GDRO (0.760).
  • Fitzpatrick Results: The method was uniquely effective at reducing bias against specific subgroups, such as non-neoplastic lesions in lighter skin tones, without sacrificing overall accuracy.
  • Disparity Reduction: In the Fitzpatrick dataset, the hybrid model achieved the lowest performance disparity ($\Delta_{nn} = 0.037$), compared to the baseline disparity of $0.083$.

Limitations and Future Directions

While promising, the approach has limitations. It remains dependent on the availability of subgroup labels during training or validation, which may not always be feasible in clinical settings. Furthermore, the reliance on a balanced validation set for fine-tuning poses challenges for rare disease applications. Future research should explore the applicability of these mechanisms to other medical modalities, such as radiology, where shortcuts may be more subtle and harder to identify.

Summary updated

BibTeX
@inproceedings{kumar2023distributionally,
  author = {Nilesh Kumar and Ruby Shrestha and Zhiyuan Li and Linwei Wang},
  title = {{Distributionally Robust Optimization and Invariant Representation Learning for Addressing Subgroup Underrepresentation: Mechanisms and Limitations}},
  booktitle = {CLIP/FAIMI/EPIMI @ MICCAI},
  year = {2023},
  doi = {10.1007/978-3-031-45249-9\_18},
}
Figure